Devin

Cognition‘s autonomous AI software engineer — an agent aimed at the work human engineers deprioritize: codebase migrations, bug backlogs, and delayed features, for customers from startups to the Fortune 500. It’s a production coding-agent product in the same category as the harnesses this spoke tracks, but pitched at long, unattended autonomy rather than an interactive terminal loop.

Long-horizon autonomy as the product

Devin’s distinguishing bet is duration: Cognition reports it running 8+ hours unattended on a real task — “through the night” — which is the long-running/durable axis pushed to a full shift. Per Cognition, that horizon is gated by the model’s ability to stay self-sufficient without losing coherence, which is why they route the work to Claude Fable 5 (earlier models drifted off-task “after minutes to an hour”). The verification posture they describe — the agent states the invariants it will maintain, executes against them, and root-causes during incident triage instead of guessing — is the loop/verification discipline made a product feature.

Trust and evaluation

Cognition’s “trust no eval” stance — validating a model on real shippable work, not leaderboards, via a proprietary “Frontier Code” anti-slop benchmark — is documented on the source page (cognition-fable5-through-the-night). It’s one of the more concrete answers the spoke has to how a serious shop actually decides a coding agent is production-ready.

Related

cognition · cognition-fable5-through-the-night · claude-fable-5 · durable-agents · loop-engineering · agentic-coding-harness